LedgerLens
A near-touchless accounts-payable pipeline: any vendor format in,
clean coded entries out, and the duplicates caught before the payment run.
Project Details
Accounts-payable staff were hand-keying thousands of documents across dozens of vendor formats. Approvals were slow, duplicate payments happened, fraud and billing errors surfaced only after the money had left, and period close dragged on for weeks. LedgerLens replaces the keying entirely. A layout-aware extraction model — OCR combined with an LLM parse — reads any document into a typed, confidence-scored schema, and handles new vendor templates without anyone writing a rule.
Automated three-way matching reconciles invoice against PO against receipt with configurable tolerance rules, so only true exceptions reach a human. Anomaly models surface duplicates, price drift and likely fraud before the payment run, not after it. Approved entries are posted, coded, to NetSuite, Tally or SAP with an immutable, audit-ready trail.
Project Research
The research problem was generalisation across vendor formats. Rule-based extraction breaks the moment a supplier redesigns their invoice, and template-per-vendor does not scale past a few dozen suppliers.
A layout-aware model that reasons about document structure rather than pixel coordinates handles a format it has never seen — which is what makes the system a platform rather than a maintenance burden.
Project Results
LedgerLens posts 94% of documents touchless at 99.5% field accuracy, closed the books 80% faster, and prevented $1.1M in leakage — duplicates, price drift and billing errors caught before payment. The AP team moved from data entry to exception handling.
